CN-121996990-A - Short-term power load prediction method integrating GIS, weather and multi-time features

Abstract

The invention belongs to the field of power system load prediction, and discloses a short-term power load prediction method integrating GIS, weather and multi-time features. The method comprises the steps of combining GIS tools to introduce spatial features such as population density and the like, constructing a load input feature set by considering meteorological features and load history features, utilizing a sparrow search algorithm to iteratively find optimal super-parameter combinations for an objective function, inputting two parallel convolution branches to synchronously extract local modes of different time scales in load data, extracting discriminative features through a multi-branch one-dimensional convolution neural network and inputting the data after feature extraction into a BIGRU network, constructing two layers of cyclic neural networks GRU in BIGRU, parallelly calculating attention weights of a plurality of subspaces based on relative position codes, introducing attention pooling layers, gradually extracting high-level features through layered nonlinear transformation by adopting a multi-layer perceptron module, and carrying out data normalization processing to give a prediction result. The method can improve the accuracy of short-term prediction results.

Inventors

• Ai Xianren
• MU YONGJUN
• Huo Zhishuo
• HU FENG
• WU JUNRU
• ZHANG JUNTAO
• CHENG CHUNTIAN
• GUAN ZHEN
• LI HONGGANG
• ZHANG QISHUN
• PENG JIE
• Li Yangyicheng
• Zhang Panquan
• ZHOU YI
• ZHAO MINGZHE

Assignees

• 华能澜沧江水电股份有限公司

Dates

Publication Date

20260508

Application Date

20260407

Claims (4)

1. 1. A short-term power load prediction method integrating GIS, weather and multi-time features is characterized by comprising the following steps: step 1, constructing an initial multisource characteristic data set; the method comprises load data acquisition, meteorological data space optimization acquisition, date characteristic construction and characteristic vector integration; Step 2, data preprocessing and normalization; carrying out normalization processing on the characteristic data obtained in the step 1, and scaling the data to a [0,1] interval by adopting a Min-Max normalization method; Step 3, optimizing super parameters based on a sparrow search algorithm; Defining model hyper-parameters and upper and lower bounds thereof, wherein the model hyper-parameters comprise training times, minimum training scale, learning rate, CNN convolution kernel number, CNN convolution kernel size, regularization parameters and hidden layer neuron number, and performing hyper-parameter optimization by adopting a sparrow search algorithm SSA and taking a mean square error MSE as an objective function; step 4, constructing a multi-scale convolution feature extraction module; Aiming at the multiscale characteristics of the power load sequence, a multiscale convolution characteristic extraction module is constructed, wherein the multiscale convolution characteristic extraction module comprises a multisection one-dimensional convolution neural network CNN and a BIGRU network, the multisource characteristic data set obtained in the step 1 is normalized in the step 2, the optimal super-parameter combination is found through a sparrow search algorithm in the step 3, and the super-parameter combination is input into the multisection one-dimensional convolution neural network CNN; The multi-branch one-dimensional convolutional neural network adopts two parallel convolutional branches, and synchronously extracts local modes with different time scales in load data, wherein the two convolutional kernels are different in size, the smaller convolutional kernel is focused on the medium-short-term fluctuation characteristics of the super-parameter combination, and the larger convolutional kernel captures medium-long-term trend changes; the method comprises the steps of realizing feature fusion of outputs of all branches through channel dimension splicing, integrating the outputs into multi-scale information through 1X 1 convolution, introducing a residual error connection mechanism, adding the original inputs to fusion features after 1X 1 convolution projection to form a cross-layer information path, extracting distinguishing features from an original load sequence through a multi-branch one-dimensional convolution neural network, digging out local dependence related to short-term load prediction from the original mixed sequence, outputting a feature sequence fused with a multi-scale local mode, inputting the feature sequence into a bidirectional gating circulation unit network BIGRU, wherein the bidirectional gating circulation unit network BIGRU comprises two layers of GRUs which are opposite to each other, a first layer BIGRU captures bidirectional context information of the original data sequence, a second layer BIGRU further extracts high-layer time sequence features, and outputting a state sequence containing the bidirectional long-range time sequence context information for short-term power load prediction by a multi-head self-attention mechanism module; Step 5, constructing a multi-head self-attention mechanism module; Introducing a multi-head self-attention mechanism based on BiGRU output, enhancing long-range dependency modeling by relative position coding, and setting the output sequence of BiGRU module as Wherein To include time steps from 1 to Each of which is a matrix of all hidden states Is one Vectors of weft are shown in A hidden state at the time; Is that Row of lines Generating a query matrix by linear transformation Key matrix Sum matrix : In the formula, 、 And Is a parameter matrix which can be learned; Representing the dimension of each attention header; For the first The attention heads have the following attention weights: in the formula, Encoding a matrix for relative position, elements thereof Representing the position And Offset by the relative distance therebetween; Ensuring that the sum of elements of each row is 1 for normalizing an exponential function; the output of the multi-head attention is obtained through splicing and linear transformation: in the formula, Representing the number of attention headers; Outputting a projection matrix; is the output of the multi-head attention mechanism; To enhance feature representation, a self-attention pooling layer is introduced: in the formula, Is the feature vector after self-attention pooling; Is the first Attention weighting coefficients for each time; is an attention weight vector; Is an attention weight matrix; Is an attention offset vector; Is a hidden dimension of the attention network; Is a hyperbolic tangent activation function; Step 6, constructing a multi-layer perceptron output module; the multi-layer perceptron MLP is designed to realize multi-step load prediction, and the forward calculation formula is as follows: in the formula, 、 、 The weight matrix is from the first layer to the third layer; 、 、 Is the corresponding offset vector; 、 、 Outputting vectors respectively from the first layer to the third layer; normalizing the layers; Activating a function for a Gaussian error linear unit, which is defined as: in the formula, As a cumulative distribution function of a standard gaussian distribution, To input scalar, equal to ; In addition, to prevent overfitting, dropout operations are applied after each layer of output: in the formula, Is the first The output after the layer Dropout, Is the first The Dropout rate of the layer is, ; For multi-step predictions, separate fully connected output heads are used: in the formula, And Respectively the first Weight vectors and bias terms for the individual prediction steps; Is the prediction step length; Is the first Output values of the prediction step sizes; The final load prediction output is: in the formula, Output vector for prediction; step 7, inversely normalizing the prediction result; After the prediction result is output by the prediction model, the prediction result is inversely normalized, and the normalization formula is as follows: in the formula, Is that Load predictors of time, i.e. Outputting a value; And Respectively the minimum value and the maximum value in the original data set in the step (2); And (5) carrying out inverse normalization on the obtained load prediction result, namely a final short-term load prediction result.
2. 2. The short-term power load prediction method integrating GIS, weather and multi-time features according to claim 1, wherein step 1 specifically comprises the following steps: step 1.1, load data acquisition, namely collecting historical power load time sequence data of a target area, wherein the unit is MW; Step 1.2, space optimization acquisition of meteorological data, which is to introduce a space characteristic layer of population density and residential area density by combining a GIS tool, determine a space range with the maximum population density through layer stacking analysis, and download meteorological data which is strongly related to load from the space range, wherein the meteorological data comprises daily average temperature, ground net radiation and ground air pressure; 1.3, constructing a date feature and a work type date feature according to a date type, wherein the date of the week is represented by 0-6 from monday to monday, and the work type date is represented by 0 and 1 for workday and holiday; and step 1.4, integrating the date factor, the weather factor and the historical load factor into a 6-dimensional input characteristic vector, wherein the characteristic vector comprises week date, working date, average temperature, ground net radiation, surface pressure and historical load time sequence data.
3. 3. The short-term power load prediction method integrating GIS, weather and multi-time features according to claim 1, wherein the normalized mathematical expression in step 2 is as follows: in the formula, Is normalized data; Is the original data; And Respectively, a maximum value and a minimum value in the data set.
4. 4. The short-term power load prediction method integrating GIS, weather and multi-time features according to claim 1, wherein the mathematical expression of sparrow location update in step 3 is as follows: discoverer location update: Follower location update: Updating the position of the alerter: in the formula, Is the number of iteration; the maximum iteration number; respectively represent the current first Sparrow of the first kind The position in the dimension is such that, Is one of a plurality of dimensions of the super parameter; Is the population scale; 、 、 the updated finder position, follower position and alerter position respectively; Is a random number; And Respectively representing an early warning value and a safety value; Is a random number subject to normal distribution; a full 1 matrix of 1×D, D being the hyper-parametric dimension; Is a1 x D matrix, each element is randomly assigned 1 or-1, and ; Is the optimal position occupied by the current discoverer; And Respectively the current global worst and optimal positions; For controlling the step length; Is a random number; is the fitness value of the current sparrow individual; And The current global optimal and worst fitness values respectively; Is a constant.

Description

Short-term power load prediction method integrating GIS, weather and multi-time features Technical Field The invention belongs to the field of power system load prediction, and relates to a short-term power load prediction method integrating GIS, weather and multi-time features. Background In recent years, the installed capacity of new energy power generation is undergoing rapid growth. However, the intermittence and fluctuation of the new energy output, the continuous running height of the power load in the superposition area, so that the power system presents a new state of 'peaceful peaked' in peaked state. During off-peak hours, abundant new energy power needs to be absorbed, while during peak hours, the maximum power load breaks through the history repeatedly, so that the power supply of partial areas tends to be tight during peak hours. In this context, the importance of a precise prediction of electrical load, especially short-term load, is unprecedented. The accurate prediction is the basis for realizing the coordination of source network and load storage, making a power generation plan, and carrying out safety check and preventive control of the power grid. The system can help a dispatching department to forward mobilize the conventional power supply, energy storage and other adjustable resources to carry out peak clipping and valley filling when the new energy output greatly fluctuates, so that clean energy is utilized to the maximum extent on the premise of guaranteeing the stable operation of a power grid, and a power supply bottom line of observe is established. The traditional load prediction model is mostly built based on influence factors such as historical load data and economic development trend, but with the aggravation of climate change, extreme weather events occur frequently, and the influence of meteorological elements on a power system is remarkably improved. Special conditions such as extreme weather events, holidays and the like not only influence the electricity utilization habit of users, but also easily cause abnormal operation of power grid equipment, and aggravate the unbalance risk of power supply and demand. Therefore, how to construct a short-term refined load prediction model has become a key problem to be solved currently. In the field of power system load prediction, the current mainstream technology is mainly divided into two main categories of traditional load prediction methods based on statistics and load prediction based on artificial intelligence methods. Traditional load prediction methods rely primarily on the time series characteristics of historical load data for modeling and prediction. As in document 1 (Shen Yang, shen Hong) a method for ultra-short-term adaptive prediction of electrical load based on the SARIMA model [ J ] the university of south-jing engineering university (natural science edition), 2024,22 (04): 78-84.) proposes an ultra-short-term adaptive electrical load prediction method based on a seasonal autoregressive moving average model, which establishes a model taking autoregressive, moving average and seasonal components into consideration by seasonal analysis and adjustment of data, and introduces a parameter adaptive strategy based on prediction errors, so as to improve the accuracy and adaptability of load prediction. However, in the context of high-proportion new energy access to a power grid, the traditional load prediction method is difficult to effectively capture the nonlinear relation between the load and uncertainty characteristics such as weather, and the prediction accuracy is difficult to guarantee. With the rapid development of computer technology and data science, the artificial intelligent methods such as machine learning, deep learning and the like are widely applied to load prediction of an electric power system. The nonlinear relationship between temperature and load is effectively captured by using a deep learning model as in literature 2（WANG Q, LU Z, et al. Modeling the impact of temperature on regional power loads using deep learning [J]. IEEE Transactions on Smart Grid, 2021, 12(4):3002-3012.）. The method for predicting the short-term load by dynamically adapting meteorological factors [ J ]. Distributed energy sources, 2024,9 (03): 73-81 ], establishes a load/meteorological information fusion module based on a parallel multi-scale time domain convolutional neural network, digs a change mode of historical load and regional weather forecast in multiple time periods, dynamically adjusts feature contribution degree and optimizes feature selection, enhances fusion of different time-space scale feature weights, and improves model prediction accuracy, as disclosed in document 3 (Deng Li, geng Lin, showei-Tong, etc.). But mainly by variable modeling, timestamp related features (e.g., seasons, months, days of the week) are ignored, the absence of which limits the model's ability to capture periodic or seasonal trends. Disclosure